30 June 2026 in SSRN Electronic Journal
Generative AI has enabled industrial-scale production of high-fidelity fake reviews that evade traditional detectors. Existing multimodal methods rely on deterministic alignment assumptions, overlooking the stochastic variability of authentic expressions and the systematic cross-modal inconsistencies inherent in AI-generated content. This study proposes the Inconsistency-Driven Adaptive Modulation Network (IDAMN), shifting detection from surface plausibility to probabilistic structural consistency verification. IDAMN employs a Mixture Density Network to model the conditional distribution of authentic multimodal data across three inconsistency dimensions—text-image semantic, text-image affective, and text-rating—quantifying deviation as probabilistic surprisal rather than deterministic distance. An inconsistency-guided FiLM mechanism dynamically modulates features by amplifying informative dissonance signals while suppressing noise without fixed thresholds. On a 1,387-sample multimodal dataset, IDAMN achieves 95.32% accuracy and 95.31% F1 score, substantially outperforming ablation variants (removing MDN drops F1 to 0.7888) and large multimodal models including Qwen-3-VL-Plus (F1=0.6512) and GLM-4.6V-106B (F1=0.6667). These findings demonstrate that explicit probabilistic inconsistency modeling provides a robust, interpretable foundation for detecting AI-generated deception, with direct implications for platform governance and trustworthy decision support in AIGC-era digital markets.